CONDUCTORS AND THE MODULI OF RESIDUAL PERFECTION JAMES M. BORGER Abstract. Let A be a complete discrete valuation ring with possibly imperfect residue field. The purpose of this paper is to give a notion of conductor for Galois representations over A which agrees with the classical Artin conductor when the residue field is perfect. The definition rests on two results of perhaps wider interest: there is a moduli space that parametrizes the ways of modifying A so that its residue field is perfect, and any Galois-theoretic object over A can be recovered from its pullback to the (residually perfect) discrete valuation ring corresponding to the generic point of this moduli space. Finally, I show that this conductor extends the non-logarithmic variant of Kato’s conductor to representations of rank greater than one. Introduction Let A be a complete discrete valuation ring of residue characteristic p > 0. If the residue field of A is perfect, there is a satisfactory theory [26, IV, VI] of ramification over A. For example, let ρ be a Galois representation over A, which is a continuous action of an absolute Galois group of the fraction field of A on a finite-rank complex (for now) vector space. Then there is a non-negative integer, the Artin conductor of ρ, that measures the extent to which ρ is ramified. If, on the other hand, we allow the residue field of A to be imperfect, ramification over A is still quite mysterious. This prevents us from understanding, say, ramification in codimension one of local systems on arithmetic surfaces. The work in this paper began with the observation that much about ramification over A can be understood by simply changing base to various extensions with perfect residue field and of relative ramification index one. The first main point (1.4) is that such extensions make up the (perfect-field-valued) points of a natural representable moduli problem. The universal residual perfection Au of A is the A-algebra corresponding to the representing object itself. It is not a discrete valuation ring—but, in a sense, only because its residue ring is not a field (and actually not even noetherian). The generic residual perfection Ag of A is the A-algebra corresponding to the fraction field of the representing object. It is a complete discrete valuation ring with perfect residue field. Both Au and Ag are, of course, unique up to unique isomorphism. It may be useful to keep a geometric analogue in mind. If we think of complete residually perfect discrete valuation rings as being like germs of curves, then it is natural to regard Au as the universal jet on A transverse to the maximal ideal and Ag as the generic jet. There is an explicit description (1.10) of these rings in the first section. For example, if A = Fp(x)[[y]], then we have u ∼ p−∞ A = Fp(¯x)[u1, u2,... ] [[y]] and g ∼ p−∞ A = Fp(¯x, u1, u2,... ) [[y]], 2 where the A-algebra structures are determined by the data x 7→ x¯+u1y+u2y +··· and y 7→ y. If A = Z\[x](p), we have u ∼ p−∞ A = W Fp(¯x)[u1, u2,... ] and g ∼ p−∞ A = W Fp(¯x, u1, u2,... ) , 1 2 J. BORGER where W denotes the functor of Witt vectors and the A-algebra structures are given by sending x to p p2 (¯x, u1, u2 ,... ). In the second section, I give some properties of Au and Ag. The most important is that the fraction field of A is algebraically closed in the fraction field of Ag. And hence the second main point: a Galois g representation ρ over A is determined by its pullback ρ|Ag to A . Therefore any invariant of ρ, such as a measure of ramification, can be recovered (in principal) from ρ|Ag . Second, I give some evidence that in defining “non-logarithmic” (or “Artin”) conductors for Galois repre- sentations over A, the most simple-minded way of proceeding along these lines is correct: the conductor ar(ρ) of ρ should be the classical Artin conductor of ρ|Ag . (Logarithmic, or “Swan”, conductors are more subtle. See below.) Taking this as the definition, we have the following result. Theorem A. Let ρ and ρ0 be Galois representations over A. (1) ar(ρ) is a non-negative integer. (2) ar(ρ ⊕ ρ0) = ar(ρ) + ar(ρ0) (3) If ρ is trivialized by a residually separable extension of A, then ar(ρ) agrees with the classical Artin conductor of ρ. (4) ar(ρ) is zero if and only if ρ is unramified. (5) ar(ρ) equals the codimension of the subspace of inertia invariants if and only if ρ is tame. This theorem is an elementary consequence of the basic properties of Ag proved in section 2. The proof of a slightly stronger version is written down in section 3. For Galois representations of rank one, Kato [17] has introduced a logarithmic conductor. As first observed in the work [20] of Matsuda (who credits it to T. Saito), this conductor has a non-logarithmic variant, which we denote arK(·). The second part of the paper is devoted to the proof of the following result. Theorem B. Let K and Kg denote the fraction fields of A and Ag. If χ is a class in H1(K, Q/Z) and χ0 1 g 0 is its image in H (K , Q/Z), then arK(χ) = arK(χ ). For present purposes, the interest in this result is the following: Corollary. Let ρ be a rank-one Galois representation over A. Then the non-logarithmic Kato conductor of ρ agrees with ar(ρ). The intuitive reason why the theorem should be true is that the order of any differential form on A should remain unchanged when the form is pulled back to Ag and that conductors of rank-one representations are essentially pole orders of differential 1-forms. When A is of equal characteristic, this has meaning and, once the necessary foundations are laid, is essentially a proof. In fact, the observation that there are residually perfect extensions with this property is what led to the definition of the general Artin conductor. In mixed characteristic, however, this provides little more than motivation, and most of this paper is spent pushing it through to a real proof. Kato’s original, logarithmic conductor is not, however, always invariant under pullback to Ag, and so the naive logarithmic analogue of ar(·) does not necessarily agree with Kato’s conductor. For some brief thoughts on logarithmic conductors for representations of higher rank, see 3.3. Of course, there is an “upper” filtration (3.5) of the absolute Galois group corresponding to this paper’s conductor, and it satisfies the Hasse-Arf property simply because residually separable extensions do. Abbes and Saito [1] also have a non-logarithmic Q -indexed upper filtration. It is tempting to hope the two >0 agree (shifted by one); a proof of this would imply the interesting result that Abbes and Saito’s filtration also satisfies the Hasse-Arf property. Boltje-Cram-Snaith [4] and Zhukov [30] also have approaches to non- abelian ramification theory; the relations with them are even more mysterious. Let me briefly describe the remaining sections. In section 4, I recall Kato’s theory, prove some basic results, and define the Kato-Artin conductor. The proof of theorem B when A is of equal characteristic is in CONDUCTORS AND THE MODULI OF RESIDUAL PERFECTION 3 section 5. It uses Matsuda’s refinement [20] of Kato’s refined Swan conductor. Because the proof in equal characteristic is so much simpler than the proof in mixed characteristic, I encourage the reader to read it first. The basic technique in mixed characteristic is to use Kato’s description [17, 4.1] (following Bloch-Kato [3]) of certain graded pieces of cohomology groups in terms of explicit K-theoretic symbols and then understand how these symbols behave under pullback to Ag. Section 7 contains a commutative diagram that encodes this behavior, and section 8 gives the proof in mixed characteristic. In the final section, I prove the corollary above. It is a pleasure to thank the Massachusetts Institute of Technology for its hospitality and the National Science Foundation for support while I was writing this paper. Conventions Throughout, p denotes a fixed prime number, and A denotes a discrete valuation ring, held fixed in each subsection, whose residue field has characteristic p. Its fraction field and residue field are denoted by K and k. All other conventions and notation are quite common, but I will state them anyway: All rings are commutative and contain 1, and all ring maps preserve 1. The fraction field of a domain R is denoted Fr(R). An extension of a field is a homomorphism to another field. An extension of a discrete valuation ring is an injective local homomorphism to another discrete valuation ring. In both cases, we usually refer to the target of the morphism, rather than the morphism itself, as the extension. We say a ring A as above is of mixed characteristic if K has characteristic 0, and is of equal characteristic if K has characteristic p. Variants of the words residue and generic refer to the residue and fraction fields of A. We also use the same words to refer to extensions. For example, we might say B/A is residually purely inseparable or is generically Galois. An extension B/A is unramified (resp. tame) if it is finite and residually sep separable and its ramification index is one (resp.